653. Harsh Aggarwal, How AI Lets Solo Consultants Take On Bigger Work

Speaker A: Hello and welcome to Unleashed. I'm your host Will Bachman and I'm here here with Harsh Agarwal who's going to tell us about Restack, which is an AI agent that works natively inside PowerPoint and takes the manual formatting off consulting and, and private equity teams. So Harsh, uh, welcome to the show.

Speaker B: Hey Will, great to have you here. Hi to everyone who's listening in. So I'm Harsh, uh, founder of Restack. As Nathal mentioned, we are building an AI agent, uh, which lives inside PowerPoint, especially for consulting team and private equity teams to take the grind of formatting off their hands and so that they can focus basically on client relationships and the actual consulting work. Just to set up the context a little bit of why we started doing this. Uh, so for decades like big firms had an edge that had nothing to do with being smarter, right. Which is production capacity. So a McKinsey team has five analysts who can turn a rough idea into a board ready deck overnight. Right. But the independent consultant has themselves at 1am doing the thinking and the slide, um, building both. Right. But AI just closed the first half of the gap in which like you have MVP tier research and synthesis with Cloud and ChatGPT and some of the other tools. But what's still open is the production layer or the production half, uh, turning that thinking into a client ready deck. And that's the last structural advantage that big firms currently have over independent consultants. And it's closing fast. And that research gap is gone. But the solo now has analytical power of an MVP pod. The production gap remains and that is why we basically were working on Restack, uh, turning that into a polished on brand deck. Silly eat hours but for solo those available hours or like sleep or some of the routine stuff, right. So and that's why we basically started in building reset. So let me show you quickly how it works. Uh, will feel free to interrupt me anytime in case uh, there's something. So let me just quickly share my screen. Okay, so if you can see it. Well basically we are, we are in normal PowerPoint, right? It's a standard deck, market research, some charts here and there, some analysis on the charts themselves. So I've kept it short, four slides, but I just wanted to show what the mechanic is. Right. So normal PowerPoint I just click and open Restack here. It opens as a sidebar inside the tool that you already live in day in, day out. And I will just ask you to do one thing right. So what I want to demonstrate is a real task that would Come up in a consulting project. But again this is just one task which can be replicated across multiple workflows and can adapt as well. So for example, uh, let me do a quick thing. So for example, I will just go

Speaker A: here and, and listeners, what I'm seeing, uh, I'll mention here we are doing a Harsh is giving sharing a screen with me. So if you're listening on audio, uh, we'll include a link in the show notes where you can go to the Permalink page for this website on the Umbrex website and we'll have some screenshots of what Harsh is showing me. We'll also have a link there where you can go and see the full video demo if you want to do that on the Umbrex website.

Speaker B: So basically I got this like new CSV, uh, file of data from one of my customer teams or client teams and I basically want to update the entire deck with this new data that has come in. Right? So let's say if I have to do it as a normal consultant, it would take my night away, right? It will take seven to eight hours updating all the deck, making sure the analysis has been updated based on the uh, updated data. But now like all these shots, I want updated. So instead of giving it to an associate or doing it myself, I just entered uh, a short prompt. Hey, take the CSV and uh, the uploaded uh, like server results and can, can you just update the deck? Right. And I'm sending this prompt right now and let's see what the agent will do. Right. So uh, here one thing uh, for the listeners that I want to highlight is in the prompt. I haven't highlighted anything about, hey, this chart should look like this or the formatting should not change. Or uh, like this color is blue or pick the data from here or there. I haven't instructed any of that, um, to the agent. Right? What I've done is the deck is open and I've just told it, hey, this is the new data. Can you just update the new data in the four slides that we have on screen? And now the agent basically goes and looks at what we have on the screen. First of all, like the four slides, it will go and read through them and it will look at the data and then automatically on its own decide, hey, this particular data needs updating on based on this. This particular storyline will change based on this. And this will go on for like a uh, couple of minutes. But then we can see the final result from here. But just to also walk, walk you through so basically, currently the agent is reading the slides and it's working on slide one, uh, at the moment, one more pointer here. So this is more like task in which we are taking, uh, data that we received from one of the internal teams or the client teams and, and updating the deck based on that. But also if it's like if I have a 10k report and I need to update my full financials based on the 10k and I need to also verify the 10k, so we can also instruct the agent for something like, hey, this is the data, but I want you to go and verify that this data is correct by going on the SEC Edgar website, downloading the 10K, verifying the data and cross referencing it and then building the slides and the data based on that. So all that is possible without any inside restack. I just wanted to show one of the workflows, uh, for your, uh, listeners and viewers.

Speaker A: So the tool is currently reading all of the slides. It, it analyzed the uploaded data and it's reading the slides and understanding all the formatted text shapes and under. So working, I guess, to understand the data and then the slide and see how they connect with one another.

Speaker B: Exactly, yes. So just to, uh, add to Will's point, uh, it will go and read the data and how it works in the backend, right? Uh, just so that your listeners can also get an understanding of how the tech works. So basically, um, if you look at, uh, the AI world in general at large, tools like Cloud code and Cursor are becoming really popular because they can augment the whole code base. And this is also what Restack does on the backend side of things. It converts the whole deck into code in the backend, so it can actually, uh, insert and modify the data wherever it's required. And that's why it can be super precise about what's changing. Instead of changing the whole deck or generating the full deck end to end, it will go and precisely check, hey, on slide 2, line number 52 needs changing from a code perspective. Um, and it will go into that particular file and just change the line or change that particular data number. And that's how it works on the backend side of things. And that tech has already been proven out, uh, with something like cursor and cloud code, uh, becoming hugely popular in the software engineering world. But that same technology can also be applied to using slides and converting the whole PowerPoint into code base and then basically letting the agent treat it as code.

Speaker A: Okay, so now we're seeing some results. It's updating the numbers and it's telling us what it's updating. So that's helpful so you can audit what happened. It's saying, okay, slide one, I changed this number 52 to 58, 41 to 45. So it, it's walking through all the changes that it made. Okay, this is slide one. Done. Slide two, done. Slide three, done. And it updated the chart and it also updated the numbers in the key takeaways on the right hand side.

Speaker B: Exactly. So uh, not only did it update the charts and some of the tables, but it will also update the storyline in that case. So for example, let's say a number goes from positive to negative in any particular instance and that uh, the storyline, what we are commending to the client, changes that will automatically get factored in as to what needs changing from that front. And right now what you're seeing Will, is it is basically taking screenshots so to see, hey. That the changes actually verified and occurred. So it verifies its own final result before presenting it to the analyst or the associate who's actually working on it.

Speaker A: Okay.

Speaker B: And just for reference, these, these are like proper charts. These are not images. So like these are completely editable. You can drag them, increase them, convert them into think cell, convert them back. So it's like just a normal, how you would work with uh, your team. It's like normal charts how you work in your consulting day to day life.

Speaker A: All right.

Speaker B: Yeah. So I think that this particular task is finished. So we uh, just, just a quick recap. We asked it to take the data and update all the data in the four slides that we have. And that's about what we asked. And the agent went through all the four slides, updated the data in all four slides. And we can see here it's converted uh, from 52 to 50, uh, 8. And if you want I can also just quickly do control Z to see how it converted it. So uh, these two slides now you can see how it got updated. So it was 44, 39 this particular slide. Let's see slide one. So it's 52 right now as you can see 52, 41, 28. And then this is the updated I did redo. So it's now updating those slides. This is a quick overview of how the change happened, what happened with the overall. So if you want I can do one more test or we can just converse about what are the implications of this. How does it work? Um, up to you.

Speaker A: Okay, well this is sort of one use case where you have updated numbers and you upload the numbers and say change any place in the chart where these numbers go. And by the way, why don't you show us what that CSV file looks like. Be helpful. Does it have sort of the before after so the program can see. Ah, ah, this was 54 and now it's changed to 58. So I'm going to change that 54 to 58. I'd just like to see what the, that CSV file looks like that you added.

Speaker B: Right. So are, uh, you able to see the CSV file on screen?

Speaker A: No. You're just sharing your PowerPoint, uh, window

Speaker B: now or still not visible?

Speaker A: No, you're just sharing your PowerPoint screen.

Speaker B: Okay.

Speaker A: You're not sharing your whole desktop.

Speaker B: No, I actually open the file inside PowerPoint, but sometimes it doesn't share. Okay, let me um, reshare uh, my screen. Okay, so this is the CSV file that I gave. Well, if you can see, doesn't have any before or after data. It just has the data that comes from the system in a way. Right. So designed for systems. So for example, what the system of, let's say the client's marketing department spits out, we just input the raw data in. So I didn't tell it to update from 52 to 58. It just figured out, hey, that is what needs updating. So it was automatically able to figure out. And in some very complex data models in which it gets, uh, it still has a choice to make that, hey, what should I update? Then it will automatically ask you sometimes that, hey, like this point is a little bit unclear. What do you want me to exactly update? It will show a pop up and then you can guide the agent according to you want it.

Speaker A: All right, great. So you showed us this one use case of I have existing slides already produced and now I have updated numbers, updated data. I want to update these slides to match the updated numbers. Uh, are there other use cases of the tool? Does it, can it just create these charts from scratch or does it, does it do other things that we should be aware of?

Speaker B: Of course. So let me share one more case in which it can create us uh, like a data model from scratch and recreate the vis visualization for it. So okay, let me just re show to the audience as well. This is basically like a normal data model. You can see it's like um, few slides. We have approximately like 5 to 10 sheets here of Excels on the screen. And I'm basically attaching this, this particular whole Excel with different screens and what I will ask it is just give me one second. All right, so basically if you can see, I attached this data model and I'm asking it, hey, can you use these acquisition numbers for an acquisition we are targeting, huh? And basically create a uh, slide about the projected user growth, right? It's like from scratch. It has to figure out, understand the data model and then create the slide. So we'll let it work and it will take like a couple of minutes and then we will see the generated slide. And uh, one interesting thing that I want to highlight here is that I didn't tell it anything about the context of the deck. Right? Um, because we are already working in this deck. So what the agent will do now, it will actually go and read the existing deck to understand, hey, like, how does this data model like even fit in with this overall uh, storyline that we have going on? Uh, what, like who are we making it for? What are the charts? Like, how is the formatting? Like, I haven't told it anything, but when we see the actual results come out, the formatting will actually match what we have on the screen already in terms of uh, the slide structure and charts and everything.

Speaker A: Now in the previous thing that we ran here, it changed all these numbers, did it not like sort of turn the whole thing to code and read it and understand it and know what it's all about when it did, did that time, does that not carry over from prompt to prompt or.

Speaker B: Uh, it does, but for example, how uh, it does is it converts it into code, but that code lives on that particular instance itself. So what I mean is, so in one particular session you can have like 10 different chats going on. So like let's say five people are working on this. They have five different chats going on. So all the context of each particular conversation is not shared with each every other conversation, but the underlying code base, which is the presentation itself is, is one. Right? So for example, let's say you and me will are working on the same deck right now and you have a, you like, your task is to create the marketing strategy and my task is to do the numbers. Let's say for instance, so you can work on the marketing strategy and the agent will still have access to the whole code base in a way. And my agent will also have access to the whole code base, but I'll be working on the financials part of it. And that is what uh, we did here. We basically started a new chat in which the previous one is already live and we can work on it. But this one basically, like, basically because this is a new workflow, the underlying code base is the same, but we are giving it a new task to do.

Speaker A: Okay, so while it's working, maybe you could summarize for us what are all the different things that the tool can do. It can. You already showed us how it can update numbers for existing slides if you give it new data. You're showing us now how it can create a new slide. Are there other things it can do? Oh, and here's the new slide.

Speaker B: Uh, okay. Yeah, so I'll just walk you through the new slide and then I will answer your question.

Speaker A: All right, sure.

Speaker B: Uh, so this is basically the projected user growth. So we asked it, can you use the acquisition numbers to create a user growth chart? And some key takeaways. So, couple of pointers here. First of all, it created everything, read the model. It was like a complex five to ten, uh, sheet model. It read, it created the charts and created some analysis on top of it. Also, uh, as I was highlighting earlier, I didn't tell it anything about the formatting. I didn't tell anything, anything about how the slide should look like or any. Who are we making it for? It just read the template, decided, hey, this is how it should look like, and automatically created it. And these charts are like native charts. So if I want to go and format it more, I can go here and change it to a different kind of formatting, kind of like you would do. Let's say you want to do some final tweaks on a chart. You can do that. And coming, uh, to your question, um, so like, you can almost think of the agent as very generalized. So whatever you would ask your associate or analyst to do, the agent can technically do that. So for example, you can ask them to research. Uh, you can ask it to research. It will go on the Internet. It might if it is linked to your companies, like basically drive or companies different kind of services. So you might have a SharePoint, you might have the internal drive, or it might get linked to a client's drive. It might be linked to different kinds of tools like PitchBook, Bloomberg or something like that. So whatever you ask it, it can pull data from any of those particular sources. And uh, whatever storyline that you want to create, it will basically go on that, create the storyline, get the data that is required from the different tools, uh, and whether it involves, uh, updating the current slides, whether it involves researching and creating new slides, whether it involves like basically creating a data model or like a financial model in the back End it will do all those things and then come back to you and say that hey, harsh or hey vo. This is what I did. How do you want me to proceed? Right, so, uh, it, uh, so to summarize, it is very generalized in nature. So uh, what depends is how much you can push it, uh, in terms of, we can try new, new and different things with it in terms of uh, you trying out a new kind of scenario, new kind of storyline. So technically it is generally capable like a normal general consultant is. But what matters more is how much more you can like navigate through it.

Speaker A: Okay. All right, now in this case it produced the chart similar to the other ones in the deck and it's formatted nicely in the same way and so forth with key takeaways on the right. Can it, could it create like any kind of chart you want, Say, oh, I want a waterfall chart showing gains and losses each month or you know, whatever. Um, can it do that?

Speaker B: Exactly. So for example, like the four slides that I showed you, uh, initially it has like different kinds of charts, right? It has column charts, bar charts. Uh, these are all created by uh, like the agent itself. I didn't create any of this. But uh, we can also try it, like we can try to convert it to a bar chart. So we can also say, let's say one more prompt in which I will say, can you. You wanted waterfall, so we'll try it. So now because we are in the same workflow and basically working on the same task, I continued the conversation and I asked it to create one more slide of the same thing of the user growth, but in a waterfall chart, uh, so that we can understand it. Although like technically it's only rising, but we'll see what it does, uh, with the waterfall. Uh, one more thing that uh, like basically um, when I give it a task, it takes like three to four minutes to complete like a normal kind of task. And that is where it is like super powerful for like independent consultants, uh, specifically because like, like for independent consulting teams because they're always short on time and they have multiple things running, managing the client, messaging, creating something for tomorrow's presentation. So the thing is that once you give it a task, you can go back to your, the other things that you are working on. So let's say you're preparing for the meeting tomorrow. You can keep compare preparing for it and it will do its job in the back end. And then your job becomes to basically review the work, which as the manager or as the lead consultant you would do any which Way. Right. Uh, for your team basically kind of does like it's an on call associate or an analyst that you always have which can review work and work alongside you rather than you increasing headcount and then basically having all those people all the time, which is not always possible for like independent consultants or like smaller firms.

Speaker A: All right, okay, so how does someone sign up to test uh, out your tool?

Speaker B: Right, so there are like multiple ways. Either you can just like go to restack.com re-stack.com it's a re hyphen stack.com or just like reach out to me on LinkedIn. I am M. Harsh Agarwal. Uh, you can reach me on LinkedIn and I will respond to you. I'll put you in a sandbox. Whoever reaches out to me just mentioned that you came from the world's podcast and ah, I will put you in a free sandbox and you can try and stress this, stress test the tool as much as you want. We are already working with some firms to stress test and uh, running some sandboxes. But I'd be happy to onboard anyone coming from the podcast and uh, giving them a free sandbox to try it out on one of the live cases that they might be working on.

Speaker A: And what's the uh, what's the pricing?

Speaker B: So it's priced uh, in like different ways. Uh, so for any individual it's priced at 1800 per month. $1800 per month?

Speaker A: 1800, correct, yes. Dollars per month, yes. Okay.

Speaker B: Uh, yes. Uh, the thing is that I sense the hesitation in your voice just to clarify the pricing as well. Basically if you look at something like PitchBook or Bloomberg, right. Firms already pay 2,500 per month for some of these tools and they don't even produce content that actually saves time for you. So for example, if you subscribe to Pitchbook or some of the other Tools, you have 2,000, 2,500 per subscription for some of these tools, which gives you data, propriety data. But again, uh, they don't really give you back the hours here. What we're doing is for example within this like 15 minute chat we were able to produce a lot of slides, right. Do a uh, serious amount of work, which it might have taken like five to six hours to just to do myself as an individual consultant. Right. And that is what the time savings is what actually matters a lot more for a consulting team, uh, because of how they price their uh, consultants. So in the end it actually saves a lot of time and money for the firm instead of uh, uh, like something like PitchBook, which does provide. Providing data, uh, or like some of the other tools that companies use. And it is sort of priced according to that particular bond.

Speaker A: Okay. So the price is $1,800 per user. Per month.

Speaker B: Per month. And for bigger teams we have 4,000 for five team. For five member teams. So it's like it comes down to 800 per seat.

Speaker A: Okay.

Speaker B: So also just to go back to the initial request that we had, so this is the waterfall of the same kind of data which is represented it uh, from tam. So it basically did the TAM SOM SAM analysis, but with the waterfall. So we have the total, active, total addressable market, segmented market. And uh, and basically it calculates the drop off in the workflow, as you can see.

Speaker A: All right. All right, fantastic. All right, well, Harsh, uh, thanks so much for being on the show today.

Speaker B: Thank you so much, Paul. Thanks everyone for joining in. Just reach out to me anytime on LinkedIn or restack.com and I'll be happy to have you in a sandbox.